Fine-Tuning, Reinforcement Learning, and Self-Learning

Speaker

Moritz Münchmeyer, University of Wisconsin, Madison

Abstract

In this talk, I will explain how reasoning models are trained using supervised fine-tuning (SFT) and reinforcement learning (RL), using our QFT fine-tuning paper (arXiv:2604.18936, MLST) as a case study. Training our own model allows us to examine how its chains of thought and distribution of errors evolve during reasoning training. In the second part, I will discuss self-learning with model-generated synthetic data. These techniques could enable models to improve without additional human-generated training examples. I will review recent academic results and present preliminary findings from our group.

Meeting Connection

Join the Lecture